| Computed tomography (CT) is one of the most common medical imaging modalities inhepatopathy diagnosis and treatment. As the first step of liver surgery assessment and planning,automatic liver segmentation from CT images plays an important role in clinical application.There are several major difficulties. First, the liver shows great anatomical variations amongstpatients; Secondly, focal liver lesions in CT images is often unpredictable. Thirdly, the intensitiesof liver and its adjacent organ like kidney and heart are very similar, which makes theirboundaries weak and fuzzy. In recent years, many researchers have proposed various techniquesfor liver segmentation. Overall, interactive and semi-automatic segmentation methods are wellstudied while automatic methods are still an open issue. Atlas-based segmentation method isvery adaptable and flexible as well, so it is a candidate approach for automatic liversegmentation. In this thesis, we present a reliable approach for automatic liver locating, andbased on that, three atlas-based methods for automatic liver segmentation are presented.The right lung in CT image was located to facilitate location of liver using the knowledge ofspatial relationship between the right lung and liver. Major transversal rotation was corrected bysymmetry in lung region. Liver lesions were extracted by combination of local thresholding anddistance knowledge. An atlas-based B-spline registration was designed to segment liver, inwhich a CT data manually delineated by expert was used as an atlas. The whole process wasproceeded without any intervention.A liver segmentation approach based on atlas demons registration was presented. Under thetheory framework of diffeomorphic demons, original demons was modified to meet the need ofliver segmentation. This approach achieved more accurate segmentation comparing to themethod based on B-spline registration.A probabilistic atlas-based liver segmentation method was proposed in which deformableregistration was constrained by integration of prior probabilistic factors. To a certain extent, thismethod allows high elastic deformation for precise segmentation while avoidingover-segmentation for stability.Experiments demonstrated that probabilistic atlas-based approach provided more accurateand robust segmentation than the single atlas-based methods, and also could yield a performancecomparable with other state-of-the-art automatic liver segmentation methods, and could be a toolfor clinical application. |
